Nested Named Entity Recognition on NEREL Russian (test)
85.81Overall Micro F1Full nested supervision
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Full nested supervisionMethod category=Upper Bound, Supervision=Full gold nested annotations2026.02 | 85.81 | 65.48 | 83.95 | |
| + Lem.Incl.+Corr.+Neutral.Method category=Weak Supervision (Fine-tuned), Technique=Lemmatized Inclusions + Corruption + Neutralization2026.02 | 77.89 | 26.37 | 82.54 | |
| + Corruption (letters, end)Method category=Weak Supervision (Fine-tuned), Technique=Entity corruption (letters, end)2026.02 | 77.81 | 25.92 | 82.39 | |
| + InclusionsMethod category=Weak Supervision (Fine-tuned), Technique=String inclusions2026.02 | 76.7 | 21.36 | 83.05 | |
| Flat (baseline)Method category=Weak Supervision (Fine-tuned)2026.02 | 76.59 | 3.84 | 83.28 | |
| Hybrid (lem. matching)Method category=LLM-based, Technique=Hybrid fine-tuned + LLM (lemmatized matching)2026.02 | 70.16 | 18.84 | 75.83 | |
| Pure LLM (DeepSeek 5-shot)Method category=LLM-based, Model=DeepSeek-R1, Few-shot setting=5-shot2026.02 | 42.39 | 5.78 | 42.63 |